sae-lens 6.14.1__py3-none-any.whl → 6.14.2__py3-none-any.whl

sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.14.1"
+__version__ = "6.14.2"
 
 import logging
 

sae_lens/evals.py

@@ -11,7 +11,7 @@ from dataclasses import dataclass, field
 from functools import partial
 from importlib.metadata import PackageNotFoundError, version
 from pathlib import Path
-from typing import Any
+from typing import Any, Iterable
 
 import einops
 import pandas as pd
@@ -24,7 +24,10 @@ from sae_lens.loading.pretrained_saes_directory import get_pretrained_saes_direc
 from sae_lens.saes.sae import SAE, SAEConfig
 from sae_lens.training.activation_scaler import ActivationScaler
 from sae_lens.training.activations_store import ActivationsStore
-from sae_lens.util import extract_stop_at_layer_from_tlens_hook_name
+from sae_lens.util import (
+    extract_stop_at_layer_from_tlens_hook_name,
+    get_special_token_ids,
+)
 
 
 def get_library_version() -> str:
@@ -109,9 +112,15 @@ def run_evals(
     activation_scaler: ActivationScaler,
     eval_config: EvalConfig = EvalConfig(),
     model_kwargs: Mapping[str, Any] = {},
-    ignore_tokens: set[int | None] = set(),
+    exclude_special_tokens: Iterable[int] | bool = True,
     verbose: bool = False,
 ) -> tuple[dict[str, Any], dict[str, Any]]:
+    ignore_tokens = None
+    if exclude_special_tokens is True:
+        ignore_tokens = list(get_special_token_ids(model.tokenizer))  # type: ignore
+    elif exclude_special_tokens:
+        ignore_tokens = list(exclude_special_tokens)
+
     hook_name = sae.cfg.metadata.hook_name
     actual_batch_size = (
         eval_config.batch_size_prompts or activation_store.store_batch_size_prompts
@@ -312,7 +321,7 @@ def get_downstream_reconstruction_metrics(
     compute_ce_loss: bool,
     n_batches: int,
     eval_batch_size_prompts: int,
-    ignore_tokens: set[int | None] = set(),
+    ignore_tokens: list[int] | None = None,
     verbose: bool = False,
 ):
     metrics_dict = {}
@@ -339,7 +348,7 @@ def get_downstream_reconstruction_metrics(
             compute_ce_loss=compute_ce_loss,
             ignore_tokens=ignore_tokens,
         ).items():
-            if len(ignore_tokens) > 0:
+            if ignore_tokens:
                 mask = torch.logical_not(
                     torch.any(
                         torch.stack(
@@ -384,7 +393,7 @@ def get_sparsity_and_variance_metrics(
     compute_featurewise_density_statistics: bool,
     eval_batch_size_prompts: int,
     model_kwargs: Mapping[str, Any],
-    ignore_tokens: set[int | None] = set(),
+    ignore_tokens: list[int] | None = None,
     verbose: bool = False,
 ) -> tuple[dict[str, Any], dict[str, Any]]:
     hook_name = sae.cfg.metadata.hook_name
@@ -426,7 +435,7 @@ def get_sparsity_and_variance_metrics(
     for _ in batch_iter:
         batch_tokens = activation_store.get_batch_tokens(eval_batch_size_prompts)
 
-        if len(ignore_tokens) > 0:
+        if ignore_tokens:
             mask = torch.logical_not(
                 torch.any(
                     torch.stack(
@@ -596,7 +605,7 @@ def get_recons_loss(
     batch_tokens: torch.Tensor,
     compute_kl: bool,
     compute_ce_loss: bool,
-    ignore_tokens: set[int | None] = set(),
+    ignore_tokens: list[int] | None = None,
     model_kwargs: Mapping[str, Any] = {},
     hook_name: str | None = None,
 ) -> dict[str, Any]:
@@ -610,7 +619,7 @@ def get_recons_loss(
         batch_tokens, return_type="both", loss_per_token=True, **model_kwargs
     )
 
-    if len(ignore_tokens) > 0:
+    if ignore_tokens:
         mask = torch.logical_not(
             torch.any(
                 torch.stack([batch_tokens == token for token in ignore_tokens], dim=0),
@@ -856,11 +865,6 @@ def multiple_evals(
                     activation_scaler=ActivationScaler(),
                     model=current_model,
                     eval_config=eval_config,
-                    ignore_tokens={
-                        current_model.tokenizer.pad_token_id,  # type: ignore
-                        current_model.tokenizer.eos_token_id,  # type: ignore
-                        current_model.tokenizer.bos_token_id,  # type: ignore
-                    },
                     verbose=verbose,
                 )
                 eval_metrics["metrics"] = scalar_metrics

sae_lens/llm_sae_training_runner.py

@@ -61,9 +61,11 @@ class LLMSaeEvaluator(Generic[T_TRAINING_SAE]):
         data_provider: DataProvider,
         activation_scaler: ActivationScaler,
     ) -> dict[str, Any]:
-        ignore_tokens = set()
+        exclude_special_tokens = False
         if self.activations_store.exclude_special_tokens is not None:
-            ignore_tokens = set(self.activations_store.exclude_special_tokens.tolist())
+            exclude_special_tokens = (
+                self.activations_store.exclude_special_tokens.tolist()
+            )
 
         eval_config = EvalConfig(
             batch_size_prompts=self.eval_batch_size_prompts,
@@ -81,7 +83,7 @@ class LLMSaeEvaluator(Generic[T_TRAINING_SAE]):
             model=self.model,
             activation_scaler=activation_scaler,
             eval_config=eval_config,
-            ignore_tokens=ignore_tokens,
+            exclude_special_tokens=exclude_special_tokens,
             model_kwargs=self.model_kwargs,
         )  # not calculating featurwise metrics here.
 

sae_lens/training/activations_store.py

@@ -29,7 +29,10 @@ from sae_lens.pretokenize_runner import get_special_token_from_cfg
 from sae_lens.saes.sae import SAE, T_SAE_CONFIG, T_TRAINING_SAE_CONFIG
 from sae_lens.tokenization_and_batching import concat_and_batch_sequences
 from sae_lens.training.mixing_buffer import mixing_buffer
-from sae_lens.util import extract_stop_at_layer_from_tlens_hook_name
+from sae_lens.util import (
+    extract_stop_at_layer_from_tlens_hook_name,
+    get_special_token_ids,
+)
 
 
 # TODO: Make an activation store config class to be consistent with the rest of the code.
@@ -113,7 +116,7 @@ class ActivationsStore:
         if exclude_special_tokens is False:
             exclude_special_tokens = None
         if exclude_special_tokens is True:
-            exclude_special_tokens = _get_special_token_ids(model.tokenizer)  # type: ignore
+            exclude_special_tokens = get_special_token_ids(model.tokenizer)  # type: ignore
         if exclude_special_tokens is not None:
             exclude_special_tokens = torch.tensor(
                 exclude_special_tokens, dtype=torch.long, device=device
@@ -763,31 +766,6 @@ def _get_model_device(model: HookedRootModule) -> torch.device:
     return next(model.parameters()).device  # type: ignore
 
 
-def _get_special_token_ids(tokenizer: PreTrainedTokenizerBase) -> list[int]:
-    """Get all special token IDs from a tokenizer."""
-    special_tokens = set()
-
-    # Get special tokens from tokenizer attributes
-    for attr in dir(tokenizer):
-        if attr.endswith("_token_id"):
-            token_id = getattr(tokenizer, attr)
-            if token_id is not None:
-                special_tokens.add(token_id)
-
-    # Get any additional special tokens from the tokenizer's special tokens map
-    if hasattr(tokenizer, "special_tokens_map"):
-        for token in tokenizer.special_tokens_map.values():
-            if isinstance(token, str):
-                token_id = tokenizer.convert_tokens_to_ids(token)  # type: ignore
-                special_tokens.add(token_id)
-            elif isinstance(token, list):
-                for t in token:
-                    token_id = tokenizer.convert_tokens_to_ids(t)  # type: ignore
-                    special_tokens.add(token_id)
-
-    return list(special_tokens)
-
-
 def _filter_buffer_acts(
     buffer: tuple[torch.Tensor, torch.Tensor | None],
     exclude_tokens: torch.Tensor | None,

sae_lens/util.py

@@ -5,6 +5,8 @@ from dataclasses import asdict, fields, is_dataclass
 from pathlib import Path
 from typing import Sequence, TypeVar
 
+from transformers import PreTrainedTokenizerBase
+
 K = TypeVar("K")
 V = TypeVar("V")
 
@@ -63,3 +65,28 @@ def path_or_tmp_dir(path: str | Path | None):
             yield Path(td)
     else:
         yield Path(path)
+
+
+def get_special_token_ids(tokenizer: PreTrainedTokenizerBase) -> list[int]:
+    """Get all special token IDs from a tokenizer."""
+    special_tokens = set()
+
+    # Get special tokens from tokenizer attributes
+    for attr in dir(tokenizer):
+        if attr.endswith("_token_id"):
+            token_id = getattr(tokenizer, attr)
+            if token_id is not None:
+                special_tokens.add(token_id)
+
+    # Get any additional special tokens from the tokenizer's special tokens map
+    if hasattr(tokenizer, "special_tokens_map"):
+        for token in tokenizer.special_tokens_map.values():
+            if isinstance(token, str):
+                token_id = tokenizer.convert_tokens_to_ids(token)  # type: ignore
+                special_tokens.add(token_id)
+            elif isinstance(token, list):
+                for t in token:
+                    token_id = tokenizer.convert_tokens_to_ids(t)  # type: ignore
+                    special_tokens.add(token_id)
+
+    return list(special_tokens)

{sae_lens-6.14.1.dist-info → sae_lens-6.14.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.14.1
+Version: 6.14.2
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE

{sae_lens-6.14.1.dist-info → sae_lens-6.14.2.dist-info}/RECORD

@@ -1,12 +1,12 @@
-sae_lens/__init__.py,sha256=bh_CgiUTwniwjnBsHPO170zHd10hLM5fCeAgMZc-8n4,3589
+sae_lens/__init__.py,sha256=U6PI8XxNzEqTakvBsTnn6i8EvoMbpcviRffWBke2frk,3589
 sae_lens/analysis/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/analysis/hooked_sae_transformer.py,sha256=vRu6JseH1lZaEeILD5bEkQEQ1wYHHDcxD-f2olKmE9Y,14275
 sae_lens/analysis/neuronpedia_integration.py,sha256=Gx1W7hUBEuMoasNcnOnZ1wmqbXDd1pSZ1nqKEya1HQc,4962
 sae_lens/cache_activations_runner.py,sha256=cNeAtp2JQ_vKbeddZVM-tcPLYyyfTWL8NDna5KQpkLI,12583
 sae_lens/config.py,sha256=IdRXSKPfYY3hwUovj-u83eep8z52gkJHII0mY0KseYY,28739
 sae_lens/constants.py,sha256=CSjmiZ-bhjQeVLyRvWxAjBokCgkfM8mnvd7-vxLIWTY,639
-sae_lens/evals.py,sha256=p4AOueeemhJXyfLx2TxOva8LXxXj63JSKe9Lnib3mHs,39623
-sae_lens/llm_sae_training_runner.py,sha256=sJTcDX1bUJJ_jZLUT88-8KUYIAPeUGoXktX68PsBqw0,15137
+sae_lens/evals.py,sha256=P0NUsJeGzYxFBiVKhbPzd72IFKY4gH40HHlEZ3jEAmg,39598
+sae_lens/llm_sae_training_runner.py,sha256=8Km519LH080RZnUBeaG2T1trq5UqxoAqokNmpX4xMTM,15200
 sae_lens/load_model.py,sha256=C8AMykctj6H7tz_xRwB06-EXj6TfW64PtSJZR5Jxn1Y,8649
 sae_lens/loading/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/loading/pretrained_sae_loaders.py,sha256=SM4aT8NM6ezYix5c2u7p72Fz2RfvTtf7gw5RdOSKXhc,49846
@@ -25,15 +25,15 @@ sae_lens/saes/transcoder.py,sha256=BfLSbTYVNZh-ruGxseZiZJ_acEL6_7QyTdfqUr0lDOg,1
 sae_lens/tokenization_and_batching.py,sha256=D_o7cXvRqhT89H3wNzoRymNALNE6eHojBWLdXOUwUGE,5438
 sae_lens/training/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 sae_lens/training/activation_scaler.py,sha256=seEE-2Qd2JMHxqgnsNWPt-DGtYGZxWPnOwCGuVNSOtI,1719
-sae_lens/training/activations_store.py,sha256=2EUY2abqpT5El3T95sypM_JRDgiKL3VeT73U9SQIFGY,32903
+sae_lens/training/activations_store.py,sha256=hHY6rW-T7sLq2a8JPEyWdm8leuIRm_MsObZs3jRTZmE,31931
 sae_lens/training/mixing_buffer.py,sha256=vDpYG5ZE70szDvBsRKcNHEES3h_WTKJ16qDYk5jPOVA,2015
 sae_lens/training/optim.py,sha256=TiI9nbffzXNsI8WjcIsqa2uheW6suxqL_KDDmWXobWI,5312
 sae_lens/training/sae_trainer.py,sha256=il4Evf-c4F3Uf2n_v-AOItCasX-uPxYTzn_sZLvLkl0,15633
 sae_lens/training/types.py,sha256=1FpLx_Doda9vZpmfm-x1e8wGBYpyhe9Kpb_JuM5nIFM,90
 sae_lens/training/upload_saes_to_huggingface.py,sha256=r_WzI1zLtGZ5TzAxuG3xa_8T09j3zXJrWd_vzPsPGkQ,4469
 sae_lens/tutorial/tsea.py,sha256=fd1am_XXsf2KMbByDapJo-2qlxduKaa62Z2qcQZ3QKU,18145
-sae_lens/util.py,sha256=lW7fBn_b8quvRYlen9PUmB7km60YhKyjmuelB1f6KzQ,2253
-sae_lens-6.14.1.dist-info/METADATA,sha256=ZE2ppvNRrI_CAr7jQ2TdcPQmfEdhLoo-UMW83KVbtvU,5318
-sae_lens-6.14.1.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
-sae_lens-6.14.1.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
-sae_lens-6.14.1.dist-info/RECORD,,
+sae_lens/util.py,sha256=tCovQ-eZa1L7thPpNDL6PGOJrIMML2yLI5e0EHCOpS8,3309
+sae_lens-6.14.2.dist-info/METADATA,sha256=WDlgsdDyQT4jmu5hxMU-pqm5PfBh0h65MTEbyuMuE3c,5318
+sae_lens-6.14.2.dist-info/WHEEL,sha256=zp0Cn7JsFoX2ATtOhtaFYIiE2rmFAD4OcMhtUki8W3U,88
+sae_lens-6.14.2.dist-info/licenses/LICENSE,sha256=DW6e-hDosiu4CfW0-imI57sV1I5f9UEslpviNQcOAKs,1069
+sae_lens-6.14.2.dist-info/RECORD,,